The price of the frontier fell twice in one day
Edition of 2026-09-23
# AI For All ### Wednesday, September 23, 2026
As of today, an operations manager who buys AI capacity by usage can point a firm's tools at the model an independent evaluator now ranks first, and budget a fifth less per unit for it, because the release that took the top spot also cut its own price.
The new leader arrived with a discount attached
Anthropic published Claude Opus 5.5 on Tuesday, the first model in a new family, and the part you feel immediately is the price list. Input and output now cost *$4/$20 per million units, a fifth below the model it replaces, and cache reads, the line item that dominates agent and coding work, are down 60%*. Anthropic says typical workloads cost forty per cent less to run, and that the model writes more than thirty per cent faster than its predecessor.
What lifts this above a routine upgrade is that the price cut and the ranking came from different places, and they agree. Artificial Analysis, an independent evaluator, places Opus 5.5 first on the index it maintains, and first on its private evaluation of professional knowledge work, ahead of every Anthropic model before it. Anthropic also raised the five-hour usage limits on its paid plans, and gave subscribers a rate limit reset they can save for when they choose.
One caution, and it is the interesting one. The two tables published about this model do not agree on how far ahead it is. Anthropic's own table reports *66.4% on Terminal-Bench 4.0, while the independent run of the same benchmark reports 59.6%*, level with a rival's flagship rather than clear of it. Different test harnesses explain part of the gap, and neither number is false. But if you are deciding what to buy on the strength of a benchmark, check whose harness produced it.
Read the announcement at Anthropic, and the independent evaluation at Artificial Analysis.
Halved prices, fewer invented answers, and two jobs that got worse
OpenAI released two models the same day, Sol and Luna, and the independent evaluation confirms the headline that matters to a buyer: prices fall by roughly half. Sol drops to *$2/$10 per million units of input and output, and Luna to $0.10/$0.50. Measured per task rather than per unit, Sol now costs $1.06 per completed evaluation task, a cut of roughly half, and Luna costs $0.07*, a cut of roughly sixty per cent.
The surprise is not the price. Both models cut their rate of confident false statements sharply, Sol from *92% to 60%* on the evaluator's knowledge benchmark. That is the kind of improvement that changes what you trust a model with. It comes with a cost, and the evaluator measures it too: Sol reaches those numbers partly by answering fewer questions, and its measured accuracy falls as a result.
Then the part that should slow you down before switching a whole workflow. In coding agents, the cheaper model regressed. On the evaluator's benchmark of economically valuable tasks, adapted from OpenAI's own dataset across dozens of occupations, both new models fell behind their predecessors. Cheaper per unit, cheaper per task, and worse on some of the work companies actually pay for.
Read the independent evaluation at Artificial Analysis.
Weak signal
Two institutions spent Tuesday making benchmark results reproducible. The UK AI Security Institute began publishing its evaluation results through Evaluation Cards, built on a schema called Every Eval Ever from a coalition of researchers, EvalEval. The first release covers five widely used benchmarks and six frontier models, with methods and configuration attached rather than summarised. Nothing here changes what a model can do today. What changes is quieter and matters more over time: the claims that decide what these models cost are becoming results somebody else can re-run.
The release comes from the UK AI Security Institute and the EvalEval Coalition, published on Hugging Face's blog.
The question of the week is not which model is best. It is whether the lower prices reach what a business pays per finished job, or stay on the sticker.